[Dave Baumann]

In the run up to Intel’s PCI Express launch NVIDIA’s "NV45" turned up, boasting PCI Express compatibility. Whilst on first glance this appeared to have been a native PCI Express implementation, closer inspection revealed NVIDIA "HSI" bridge chip integrated onto the graphics core’s package. Here we’ll take a look at NV45, in its GeForce 6800 GT PCI Express guise, and measure its performance against the standard AGP GeForce 6800.

"NVIDIA's PCX line of boards, which featured AGP GeForce FX's bridged with the HSI chip, all had very different layouts from their standard AGP counterparts, which was required due to the discrete HSI chip on the board which resulted in the core and memory needed to be shifted around the board. With the high memory speeds and 256-bit bus of GeForce 6800, board space and positioning is at a premium as the trace paths and lengths are fairly critical, hence a discrete HSI solution may have been difficult to implement."

Read the full article here.

[Hanners]

Spotted one typo under the board specifications:

500MHz GDDR-3 (1.1GHz Effective)

Now, either my maths is off, or 500MHz DDR is 1GHz effective.

[ET]

Regarding the Doom benchmark, this was a good opportunity to test the maximum quality, that requires 512MB on the card.

[Dave Baumann]

Originally Posted by ET

Regarding the Doom benchmark, this was a good opportunity to test the maximum quality, that requires 512MB on the card.

Although I didn't show it, I did test it and it appears to make bugger all difference - there is a few FPS performance hit on both the AGP and PCIe 6800 GT, but nothing particularly significant. I have a PCIe X800 XT back now so I might do a comparison of the two a little later on - I might also ask around a bit to see if there is any explaination of why there is such little performance differences between AGP and PCIe (although, I don't yet actually know if the X800 XT shows performance improvements relative to an AGP X800 XT).

Andy - Thanks, updated.

[Anonymous]

Opening page: "Teir-1" -> "Tier-1"

Good article!

Side note: I couldn't remeber what "AIB" is ("OEM" I remember) and couldn't be arsed to google it up, but then, I probably didn't need to know. ;-)

[Tokelil]

A very good article IMO. Putting the AMD system in there was a really good way to see which wins was due to the system rather than the bus.

[Richard]

Good article. Noticed a typo with a "5800 GT" but otherwise very informative. I'm surprised the Ultra setting on D3 didn't make a (significant) dent on AGP.

Btw, Dave, have you ever typed "r_renderer R200" on any ATI card that usually runs D3 in ARB2? Toggling between the two paths might surprise you how much the overall rendering quality drops and might change your mind that the general light shader "only" takes advantage of Dx8.1-level hw.

[Neeyik]

Originally Posted by Guest367863

Side note: I couldn't remeber what "AIB" is ("OEM" I remember) and couldn't be arsed to google it up, but then, I probably didn't need to know.

AIB = Add In Board (eg. GeCube, Asus, Hercules are all AIB partners).

[MY80S]

Since the HSI is integrated in nv45, shouldn't the bus standard read peg instead of agp8x in the chip specs table ?

Nice article though

[Rodéric]

In the conclusion

As we can see from this test, NVIDIA's HSI Bridge solution if fairly flexible in terms is usage and allows them to turn around boards ready to be utilised...

I fail to understand the bold part, I think there's at least a typo.

Interesting article overall, any way to know if it's a driver issue or being not native that makes the card crawl when reading RAM ?

[Dave Baumann]

Originally Posted by Mordenkainen

Btw, Dave, have you ever typed "r_renderer R200" on any ATI card that usually runs D3 in ARB2? Toggling between the two paths might surprise you how much the overall rendering quality drops and might change your mind that the general light shader "only" takes advantage of Dx8.1-level hw.

Originally Posted by 6800 GT Preview

Doom3's primary rendering characteristic, being its unified lighting model, requires plenty of stencil fill-rate and a lighting shader that can be achieved in a single pass per light in DirectX8.1 equivalent hardware.

Opinion aside (although I wasn't typing my opinion) has anything in Doom3 changed that no longer allows the lighting shader to operate in one pass on PS1.4 class hardware?

Originally Posted by MY80S

Since the HSI is integrated in nv45, shouldn't the bus standard read peg instead of agp8x in the chip specs table ?

Its only integrated on to the package, not into the chip - the graphics core itself is still AGP8X. You might argue that this chip is AGP16X, although technically its overclocked AGP8X, rather than what you might expect an AGP16X spec to be, if it were to follow the previous AGP upgrade path.

Originally Posted by Ingenu

Interesting article overall, any way to know if it's a driver issue or being not native that makes the card crawl when reading RAM ?

Its possible, although I would say that the Serious Magic Image Uploading tests shows the same trend as TechReport's testing, and that was done with earlier drivers.

[WaltC]

Originally Posted by DaveBaumann

Its only integrated on to the package, not into the chip - the graphics core itself is still AGP8X. You might argue that this chip is AGP16X, although technically its overclocked AGP8X, rather than what you might expect an AGP16X spec to be, if it were to follow the previous AGP upgrade path.

Although the PCIe bus picture is still fairly murky, this does at least *seem* to explain why the ATi position relative to bridging an AGP chip to PCIe is the inverse of nVidia's. While on paper the "AGP x16" bridge route might appear the more economical approach during the AGP to PCIe transition (the length of which is anybody's guess), if there are problems relative to the "overhead" required by the bridge integration such that performance expectations are greatly diminished for it in relation to either "AGP x16" or native PCIex16, then the bridge-chip approach might well turn out to be far more expensive in terms of the PCIe sales it costs nVidia. Perhaps by the time nVidia's ready to sell them and system OEMs are ready to buy them nVidia will have come up with a native PCIe version.

[Anonymous]

Dave, I appreciate the work that you've done and the time that you've invested in testing all 3 rigs with a wide variety of software!

However, don't you think that the 10 pages essentially showing that there's no performance difference between NV40/GT and NV45/GT could've been cut down to 2 or 3 wtih commentary that this behaviour is exemplary for all apps you tested?

[Neeyik]

The preview contains 2/3 pages of commentary...plus 7/8 pages of proof. That's one way of looking at it.

[Dave Baumann]

Originally Posted by WaltC

Perhaps by the time nVidia's ready to sell them and system OEMs are ready to buy them nVidia will have come up with a native PCIe version.

Well, they already have - 6600.

Originally Posted by incurable_nli

However, don't you think that the 10 pages essentially showing that there's no performance difference between NV40/GT and NV45/GT could've been cut down to 2 or 3 wtih commentary that this behaviour is exemplary for all apps you tested?

That depends on how you want to look at things. You may want to look at the article purely from the perspective of the difference between AGP and PCIe 6800 GT, or you may want to look at it as this and an article on how the 6800 GT might perform in games - not everyone reading it will have read our other reviews.

However, having taken the numbers its not a huge amount more to include them - so people who want to know the performance of the GT in general or people who want to know how it performs on different platforms will get something out of this. I, of course, also get a bunch of extra page clicks (we are, of course, not forcing you through the pages as there is a skip menu available)

[Richard]

Originally Posted by DaveBaumann

Opinion aside (although I wasn't typing my opinion) has anything in Doom3 changed that no longer allows the lighting shader to operate in one pass on PS1.4 class hardware?

AFAIK, the r200 is still single pass. That does not mean, however, that if you have a card capable of running in the ARB2 path it will look like in R200. The difference between r200 and arb2 is quite significative so I was just wondering why you mentioned ps1.4 and single pass. It might give the idea that anything above ps 1.4 will not give people a better overall lighting shader.

[Moloch]

What was the issue with the R200 and Doom3, where JC said the R200 should be running faster than the GF4, but it isn't?
Sorry about OT.

[Anonymous]

Isn't there already an sli connection with the board? Or what is that connector for example here.

[Anonymous]

Originally Posted by incurable_nli

However, don't you think that the 10 pages essentially showing that there's no performance difference between NV40/GT and NV45/GT could've been cut down to 2 or 3 wtih commentary that this behaviour is exemplary for all apps you tested?

Dave's reasoning aside, I'd like to add that b3d's extensive proofs and consistent, attractive review layouts are big reasons I keep coming back. Thanks again!

[Dave Baumann]

Originally Posted by Slovo

Dave's reasoning aside, I'd like to add that b3d's extensive proofs and consistent, attractive review layouts are big reasons I keep coming back. Thanks again!

The review layout and the consistency is something that I had hoped would breed familiarity, however even I will questions your seriousness of that statement with extensive proofing!!!

[2senile]

Originally Posted by DaveBaumann

extensive proofing!!!

Use a damned spell checker for Fs sake "Wavey"!

[Anonymous]

Well, I hope the PCI Express version of this board will be compatible with the future.

It's one of the reasons I'm considering on buying it.

Considering that, where can I find this board? Anyone know where or when?

Thanks

[Dave Baumann]

Limited availablity appears to be creeping out right now - it probably depends on where you are based.

[Anonymous]

I've approached: Gigabyte, MSI, ASUS and ELSA and I can't find them making this monster.
Thanks for your reply. Do you know if any OEM maker has announced anything?

[Dave Baumann]

Apologies - I misread this as one of the 6600 threads!

Sorry, no, ther has been no information as to availablility on the PCIe version of 6800 GT. The closest I've seen so far is the Dell option of a "6800 GTO", however this appears to be 12 pipelines with reduced clocks and still two slots. I've not heard of any others (although it may be worthwhile digging around the likes of Alienware, Falcon Northwest, etc. to see if they are offering anything).